AI SECURITY
Small model. Fast decision.
Before the expensive model
ever sees it.
Security bolted onto AI after deployment will lose. Most AI security today consists of dashboards that monitor activity and alert after the damage is done. A prompt injection has already exfiltrated data. A jailbroken tool call has already run. The alert is a post-mortem, not a defense.
Security needs to participate in the request, decision, tool and execution loops — before execution. Haiku Guard evaluates requests and actions before the model ever sees them.
It runs at the edge of inference: a small, fast classifier that answers one question — is this safe to proceed?
Every request and every tool action is classified before it proceeds. Known attack patterns are blocked with high confidence; policy-sensitive actions are flagged for review.
Don't ask a storyteller to be your firewall.
Most security decisions don't need generation. They need classification. And they need it quickly — before the expensive model, before the tool, before the side effect. Haiku Guard runs at the edge of inference, adding milliseconds, not seconds, to every request.
Haiku Guard is a classification-based security layer. It can detect known attack patterns with high confidence, but novel attack vectors may evade classification. It is designed as part of a defense-in-depth strategy — paired with Sokka for control and FireAgent for isolation — not a replacement for comprehensive security practices.